It is tempting to treat an AI summary as a finished product. Paste in a paper, get a paragraph, ship it. But a paragraph is not a research artifact, and the difference is not cosmetic.
Summaries detach claims from evidence
A summary's job is compression, and compression drops the connective tissue — the figure a claim rests on, the method that produced it, the caveat that bounds it. What's left can read as more certain than the research itself.
Artifacts preserve context
A research artifact keeps the structure: claims linked to evidence, methods explained, limitations surfaced, data connected. It supports multiple depths of reading rather than collapsing everything to one.
A bad summary says "this paper proves X." A good artifact says "this paper presents evidence for X under these conditions, with these limitations — here's how to inspect it."
Authors need control
An artifact also assumes review. AI can draft, organize, and explain, but researchers should be able to correct, approve, and lock the public language. That human step is what separates assistance from automation.
Research needs reusable formats
Finally, an artifact is structured data, not just rendered text — which means it can be re-themed, exported, and updated over time. A summary is a dead end. An artifact is infrastructure.